Abstract
Spatial modeling is one of the growing areas of research in economics in recent years. However, these models are not tested enough. Even if tests are performed, they are done in a piece-wise fashion. Another age-long problem in economic modeling is endogeneity of one or more variables. Endgeneity is caused due to a number of reasons one of which is simultaneous modeling of economic variables. This paper considers specification testing in the context of a spatial autoregresive (SAR) model with an endogenous regressor. First, we construct standard Rao’s Score (RS) tests for null hypothesis of the absence of spatial autocorrelation and endogeneity. These standard RS tests are invalid in the presence of local misspecification of the models under the alternative hypotheses. Therefore, in our next step, we develop adjusted tests using the technique of Bera and Yoon (Econom Theor 9:649–658, 1993), that are robust to local misspecification. These adjusted (or robustified) tests are simple to calculate and easy to implement. With a Monte Carlo study we investigate the finite sample performance of all the proposed tests, and the results confirm that the robust tests perform better compared to their non-robust counterparts both in terms of size and power.
Similar content being viewed by others
Notes
We are thankful to a Referee for suggesting us to include a discussion on the impact of W (sparse or dense) on the test statistics.
We appreciate this suggestion from a Referee.
References
Anselin L (2010) Thirty years of spatial econometrics. Pap Reg Sci 89:3–25
Anselin L (2013) Spatial econometrics: methods and models. Springer, Berlin
Anselin L, Bera AK (1998) Spatial dependence in linear regression models with an introduction to spatial econometrics. Handb. Appl. Econ. Stat. 155:237–290
Anselin L, Bera AK, Florax R, Yoon MJ (1996) Simple diagnostic tests for spatial dependence. Reg Sci Urban Econ 26:77–104
Anselin L, Cohen J, Cook D, Gorr W, Tita G (2000) Spatial analyses of crime. Crim Justice 4:213–262
Arbia G, Ghiringhelli C, Mira A (2019) Estimation of spatial econometric linear models with large datasets: How big can spatial big data be? Reg Sci Urban Econ 76:67–73
Bera AK, Yoon MJ (1993) Specification testing with locally misspecified alternatives. Econom Theor 9:649–658
Bera AK, Montes-Rojas G, Sosa-Escudero W (2009) Testing under local misspecification and artificial regressions. Econ Lett 104:66–68
Bera AK, Bilias Y, Yoon MJ, Taşpınar S, Doğan O (2020) Adjustments of Rao’s Score Test for distributional and local parametric misspecifications. J Econom Methods 9:1–29
Cheng W, Lee LF (2017) Testing endogeneity of spatial and social networks. Reg Sci Urban Econ 64:81–97
Cliff AD, Ord JK (1973) Spatial autocorrelation. Pion, London
Davidson R, MacKinnon JG (1987) Implicit alternatives and the local power of test statistics. Econometrica 55:1305–1329
Drukker DM, Egger P, Prucha IR (2013) On two-step estimation of a spatial autoregressive model with autoregressive disturbances and endogenous regressors. Econom Rev 32:686–733
Fang Y, Park SY, Zhang J (2014) A simple spatial dependence test robust to local and distributional misspecifications. Econ Lett 124:203–206
Jenish N, Prucha IR (2009) Central limit theorems and uniform laws of large numbers for arrays of random fields. J Econom 150:86–98
Jenish N, Prucha IR (2012) On spatial processes and asymptotic inference under near-epoch dependence. J Econom 170:178–190
Kelejian HH, Prucha IR (2004) Estimation of simultaneous systems of spatially interrelated cross sectional equations. J Econom 118:27–50
Lee L-F (2002) Consistency and efficiency of least squares estimation for mixed regressive, spatial autoregressive models. Econom Theor 18:252–277
Lee L-F, Yu J (2009) Spatial nonstationarity and spurious regression: the case with a row-normalized spatial weights matrix. Spat Econ Anal 4:301–327
Liu X (2012) On the consistency of the LIML estimator of a spatial autoregressive model with many instruments. Econ Lett 116:472–475
Liu X, Lee LF (2013) Two-stage least squares estimation of spatial autoregressive models with endogenous regressors and many instruments. Econom Rev 32:734–753
Liu X, Saraiva P (2015) GMM estimation of SAR models with endogenous regressors. Reg Sci Urban Econ 55:68–79
Pace RK, Barry R (1997) Sparse spatial autoregressions. Stat Probab Lett 33:291–297
Qu X, Lee LF (2015) Estimating a spatial autoregressive model with an endogenous spatial weight matrix. J Econom 184:209–232
Saikkonen P (1989) Asymptotic relative efficiency of the classical test statistics under misspecification. J Econom 42:351–369
Smith TE (2009) Estimation bias in spatial models with strongly connected weight matrices. Geogr Anal 41:307–332
Stakhovych S, Bijmolt TH (2009) Specification of spatial models: a simulation study on weights matrices. Pap Reg Sci 88:389–408
Villar OA (1999) Spatial distribution of production and international trade: a note. Reg Sci Urban Econ 29:371–380
Wang X, Guan J (2017) Financial inclusion: measurement, spatial effects and influencing factors. Appl Econ 49:1751–1762
Whittle P (1954) On stationary processes in the plane. Biometrika 41:434–449
Yang K, Lee LF (2017) Identification and QML estimation of multivariate and simultaneous equations spatial autoregressive models. J Econom 196:196–214
Acknowledgements
We are most grateful to two anonymous referees for their pertinent comments and helpful suggestions which greatly improved the content and exposition of the paper. An earlier version of this paper was presented at the 19th International Workshop “Spatial Econometrics and Statistics,” Nantes, France, May 31 to June 1, 2021. We are most grateful to the participants of that workshop for comments, especially to the Discussant of our paper Professor Anna Gloria Billé for careful reading of the draft version of the paper and suggesting many pertinent comments and improvements that we have incorporated. Thanks are also due to Professors Osman Doğan and Süleyman Taşpınar for their helpful suggestions. Of course, we retain the responsibility for any remaining errors.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix 1: Score functions and the Hessian matrix
Using the expression of the log-likelihood function in Eq. (4.1), here we derive the first- and the second-order derivatives.
1.1 Scores
where \(G=S^{-1}W\) and \(S=(I_{n}-\lambda W)\). These score functions evaluated under the joint null \(H_{0}:\lambda _{0}=0,\,\delta _{0}=0\), give us the scores in Eq. (4.2) that were used in developing the test statistics in Sect. 4.
Taking derivatives of the above score functions we obtain the second-order derivatives, i.e., the elements of the Hessian matrix, as follows:
1.2 Elements of the Hessian matrix
Appendix 2: Proofs of proposition and corollaries
1.1 Proof of Proposition 1
We start with the first result. To do so we derive the asymptotic normality of \(d_{\alpha _{2}}(\tilde{\theta })\) under \(H_{A}^{\alpha _{2}}:\alpha _{20}=\alpha ^{*}_{2}+\frac{\tau _{2}}{\sqrt{n}}\), and \(H_{A}^{\alpha _{3}}:\alpha _{30}=\alpha ^{*}_{3}+\frac{\tau _{3}}{\sqrt{n}}\). Let \(\theta =(\alpha _{1}^{'},\,\alpha _{2}^{'},\,\alpha _{3}^{'})^{'}\) be an arbitrary value of \(\theta\) and \(\tilde{\theta }=(\tilde{\alpha _{1}}^{'},\,\alpha ^{*'}_{2},\,\alpha ^{*'}_{3})^{'}\) be the restricted MLE of \(\theta\) under the joint null. Let us expand the score function w.r.t. \(\alpha _{2}\), i.e., \(\frac{\partial l(\tilde{\theta })}{\partial \alpha _{2}}\), around \(\theta _{0}=(\alpha _{10}^{'},\,\alpha _{20}^{'},\,\alpha _{30}^{'})^{'}\) using Taylor series expansion as
where \(``\overset{a}{=}"\) means asymptotic equivalence.
Since \((\alpha ^{*}_{2}-\alpha _{20})=-\frac{\tau _{2}}{\sqrt{n}}\) and \((\alpha ^{*}_{3}-\alpha _{30})=-\frac{\tau _{3}}{\sqrt{n}}\) under \(H_{0}\), we have
Now expanding \(\frac{\partial l(\tilde{\theta )}}{\partial \alpha _{1}}\) around \(\theta _{0}=(\alpha _{10}^{'},\,\alpha _{20}^{'},\,\alpha _{30}^{'})^{'}\), we obtain
Since \(\frac{\partial l(\tilde{\theta )}}{\partial \alpha _{1}}=0\), where \(\tilde{\theta }=(\tilde{\alpha _{1}}^{'},\,\alpha ^{*'}_{2},\,\alpha ^{*'}_{3})^{'}\) is the restricted MLE of \(\theta\), we have from the above equation
Substituting the expression of \(\sqrt{n}(\tilde{\alpha _{1}}-\alpha _{10})\) from (4.17) in Eq. (4.16) we obtain
Denoting \(J_{\alpha _{2}\cdot \alpha _{1}}\equiv J_{\alpha _{2}\alpha _{2}} - J_{\alpha _{2}\alpha _{1}}J^{-1}_{\alpha _{1}\alpha _{1}}J_{\alpha _{1}\alpha _{2}}\) and \(J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}} \equiv J_{\alpha _{2}\alpha _{3}}-J_{\alpha _{2}\alpha _{1}}J^{-1}_{\alpha _{1}\alpha _{1}}J_{\alpha _{1}\alpha _{2}}\), from the above equation we have
where \(R=\frac{1}{\sqrt{n}}\frac{\partial l(\theta _{0})}{\partial \alpha _{2}}-J_{\alpha _{2}\alpha _{1}}J^{-1}_{\alpha _{1}\alpha _{1}}\frac{1}{\sqrt{n}}\frac{\partial l(\theta _{0})}{\partial \alpha _{1}}\) is a random variable with mean zero while \(C=J_{\alpha _{2}\cdot \alpha _{1}}\tau _{2} + J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}\tau _{3}\) is simply a constant.
Therefore,
where \(V=Var(R)=(J_{\alpha _{2}\alpha _{2}}-J_{\alpha _{2}\alpha _{1}}J^{-1}_{\alpha _{1}\alpha _{1}}J_{\alpha _{1}\alpha _{2}})=J_{\alpha _{2}\cdot \alpha _{1}}\). Thus, we can write
The asymptotic distribution of \(RS_{\alpha _{2}}\) can be easily obtained using the asymptotic distribution of \(d_{\alpha _{2}}\) in Eq. (4.18) as follows:
where \(\nu _{3}\equiv \nu _{3}(\tau _{2},\,\tau _{3})=(J_{\alpha _{2}\cdot \alpha _{1}}\tau _{2} + J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}\tau _{3})^{'}J^{-1}_{\alpha _{2}\cdot \alpha _{1}}(J_{\alpha _{2}\cdot \alpha _{1}}\tau _{2} + J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}\tau _{3})=\tau _{2}^{'}J_{\alpha _{2}\cdot \alpha _{1}}\tau _{2} + 2\tau _{2}^{'}J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}\tau _{3} + \tau _{3}^{'}J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}J^{-1}_{\alpha _{2}\cdot \alpha _{1}}J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}\tau _{3}\).
This proves the first part of Proposition 1. The second part of the proposition states:
Under \(H_{o}^{\alpha _{2}}:\alpha _{20}=\alpha ^{*}_{2}\) and irrespective of the value of \(\alpha _{3}\) we have
To do this let us define \(\kappa =(\alpha _{2}^{'},\,\alpha _{3}^{'})^{'}\) and \(l_{\kappa }(\tilde{\theta })=\frac{\partial l(\tilde{\theta )}}{\partial \kappa }=\big [l^{'}_{\alpha _{2}}(\tilde{\theta }),\,l^{'}_{\alpha _{3}}(\tilde{\theta })\big ]^{'}\), say.
Expanding \(l_{\kappa }(\tilde{\theta })\) around \(\theta _{0}=(\alpha ^{'}_{10},\,\kappa ^{'}_{0})^{'}\) using first-order Taylor series expansion, we get
where D denotes the second order derivative matrix, namely
Similarly, expanding \(l_{\alpha _{1}}(\tilde{\theta })\) by Taylor series expansion we obtain
Combining Eqs. (4.20) and (4.21), we get
where \(J_{\kappa \alpha _{1}}=\begin{pmatrix} J_{\alpha _{2}\alpha _{1}}\\ J_{\alpha _{3}\alpha _{1}} \end{pmatrix}\). Substituting \(\tau _{2}=0\) in Eq. (4.22) gives us distribution of \(\sqrt{n}d_{\kappa }(\tilde{\theta })\) under \(H^{\alpha _{2}}_{0}:\alpha _{20}=\alpha ^{*}_{2}\) and \(H^{\alpha _{3}}_{A}:\alpha _{30}=\alpha ^{*}_{3}+\tau _{3}/\sqrt{n}\) as
implying
The adjusted score w.r.t. \(\alpha _{2}\) can be written as
The distribution of the adjusted score is obtained from the distribution in Eq. (4.23) as
Thus, for \(RS^{*}_{\alpha _{2}}\) which is based on \(d^{*}_{\alpha _{2}}\), we have the following distribution under \(H^{\alpha _{2}}_{0}:\alpha _{20}=\alpha ^{*}_{2}\) and irrespective of the value of \(\alpha _{3}\)
Finally, the last part of Proposition 1 states:
Under \(H^{\alpha _{2}}_{A}:\alpha _{20}=\alpha ^{*}_{2}+\tau _{2}/\sqrt{n}\) and irrespective of the value of \(\alpha _{3}\)
where \(\nu _{4}\equiv \nu _{4}(\tau _{2})=\tau _{2}^{'}J_{\alpha _{2}\cdot \alpha _{1}}\tau _{2}-\tau _{2}^{'}J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}J^{-1}_{\alpha _{3}\cdot \alpha _{1}}J_{\alpha _{3}\alpha _{2}\cdot \alpha _{1}}\tau _{2}\) is the non-centrality parameter. To prove this, note that similarly as in (4.23), the distribution of \(\sqrt{n}d_{\kappa }(\tilde{\theta })\) under \(H^{\alpha _{2}}_{A}\) and \(H^{\alpha _{3}}_{0}\) can be obtained by substituting \(\tau _{3}=0\) in (4.22) as
Thus, the distribution of \(d^{*}_{\alpha _{2}}\) can be obtained from (4.24) as
which implies
where \(\nu _{4}\equiv \nu _{4}(\tau _{2})=\tau _{2}^{'}(J_{\alpha _{2}\cdot \alpha _{1}}-J_{\alpha _{2}\alpha _{3}\cdot \alpha _{1}}J^{-1}_{\alpha _{3}\cdot \alpha _{1}}J_{\alpha _{3}\alpha _{2}\cdot \alpha _{1}})\tau _{2}\).
1.2 Proof of Corollaries
In this section we derive the non centrality parameters of the tests in the context of our model as given in Corollaries 2 and 3.
Proof of Corollary 2
The first part of Corollary 2 states that
Under \(H^{\lambda }_{A}:\lambda _{0}=\frac{\tau _{2}}{\sqrt{n}}\) and \(H^{\delta }_{A}:\delta _{0}=\frac{\tau _{3}}{\sqrt{n}}\)
where \(\nu _{9}\equiv \nu _{9}(\tau _{2},\,\tau _{3})=\tau ^{2}_{2}J_{\lambda \cdot \psi } + 2\tau _{2}\tau _{3}J_{\lambda \delta \cdot \psi } + \tau ^{2}_{3}\frac{J^{2}_{\lambda \delta \cdot \psi }}{J_{\lambda \cdot \psi }}\). Comparing with our general set-up in Sect. 3, here we have \(\alpha _{2}\equiv \lambda\), \(\alpha _{3}\equiv \delta\) and the set of nuisance parameters as \(\alpha _{1}\equiv \psi =(\phi ,\,\beta ^{'},\,\gamma ^{'},\,\sigma _{\xi }^{2},\,\sigma _{2}^{2})^{'}\). Therefore, substituting these in \(\nu _{3}\) of Proposition 1 we get
The second part of Corollary 2 states
Under \(H^{\lambda }_{A}:\lambda _{0}=\frac{\tau _{2}}{\sqrt{n}}\)
where \(\nu _{10}\equiv \nu _{10}(\tau _{2})=\tau ^{2}_{2}[J_{\lambda \cdot \psi }-\frac{J^{2}_{\lambda \delta \cdot \psi }}{J_{\delta \cdot \psi }}]\).
The proof for the non-centrality parameter directly follows from \(\nu _{4}\) in the last part of Proposition 1.
\(\square\)
Proof of Corollary 3
Similarly as in Corollary 2, we obtain the non centrality parameters in Corollary 3 by substituting \(\alpha _{2}\equiv \delta\) and \(\alpha _{3}\equiv \lambda\) and \(\alpha _{1}\equiv \psi =(\phi ,\,\beta ^{'},\,\gamma ^{'},\,\sigma _{\xi }^{2},\,\sigma _{2}^{2})^{'}\).
The first part of Corollary 3 says
Under \(H^{\delta }_{A}:\delta _{0}=\frac{\tau _{3}}{\sqrt{n}}\) and \(H^{\lambda }_{A}:\lambda _{0}=\frac{\tau _{2}}{\sqrt{n}}\)
where \(\nu _{11}\equiv \nu _{11}(\tau _{2},\,\tau _{3})=\tau ^{2}_{3}J_{\delta \cdot \psi } + 2\tau _{2}\tau _{3}J_{\lambda \delta \cdot \psi } + \tau ^{2}_{2}\frac{J^{2}_{\lambda \delta \cdot \psi }}{J_{\delta \cdot \psi }}\).
Following Proposition 1 we have
The second part of Corollary 3 states
Under \(H^{\delta }_{A}:\delta =\frac{\tau _{3}}{\sqrt{n}}\) we have
where \(\nu _{12}\equiv \nu _{12}(\tau _{3})=\tau ^{2}_{3}[J_{\delta \cdot \psi }-\frac{J^{2}_{\lambda \delta \cdot \psi }}{J_{\lambda \cdot \psi }}]\).
Again from Proposition 1 we directly obtain
\(\square\)
Rights and permissions
Springer Nature or its licensor holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Koley, M., Bera, A.K. Testing for spatial dependence in a spatial autoregressive (SAR) model in the presence of endogenous regressors. J Spat Econometrics 3, 11 (2022). https://doi.org/10.1007/s43071-022-00026-7
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s43071-022-00026-7